CVSROOT: /cvs
Module name: src
Changes by: [email protected] 2025/11/17 22:24:38
Modified files:
sys/arch/powerpc64/include: atomic.h
Log message:
Change powerpc64 memory barriers to "sync"
The switch from "isync" to "sync" fixes the parking mutex on a 4-core
IBM POWER9, and proves that "isync" is the wrong barrier.
Back in June 2024, I suspected that "isync" was wrong, gave an
explanation, and proposed a diff switching to "sync" in
https://marc.info/?l=openbsd-tech&m=171816471216159&w=2
I forgot the diff, and lacked evidence of a problem until this month
(November 2025), when the parking mutex in /sys/kern/kern_lock.c
caused lockups. The 4-core POWER9 froze in "make -j4 build" before
finishing libc, but the switch from "isync" to "sync" fixes it.
ok jca@ dlg@
